Computer Speech & Language 63 (2020) 101075
Contents lists available at ScienceDirect
Computer Speech & Language
journal homepage: www.elsevier.com/locate/csl
Multilingual stance detection in social media political debates
Mirko Laia,1,*, Alessandra Teresa Cignarellaa,b,1, Delia Irazu Hernandez Faríasc, Cristina Boscoa, Viviana Pattia, Paolo Rossob a Dipartimento di Informatica, Universita degli Studi di Torino, Italy b PRHLT Research Center, Universitat Polit ecnica de Val encia, Spain c Division de Ciencias e Ingenierías, Universidad de Guanajuato Campus Leon, Mexico
ARTICLE INFO ABSTRACT
Article History: Stance Detection is the task of automatically determining whether the author of a text Received 18 January 2018 is in favor, against, or neutral towards a given target. In this paper we investigate the Revised 30 December 2019 portability of tools performing this task across different languages, by analyzing the Accepted 4 February 2020 results achieved by a Stance Detection system (i.e. MultiTACOS) trained and tested in a Available online 12 February 2020 multilingual setting. Firstofall,asetofresourceson topics related to politics for English, French, Italian, Keywords: Spanish and Catalan is provided which includes: novel corpora collected for the pur- Stance detection Multilingual pose of this study, and benchmark corpora exploited in Stance Detection tasks and Contextual features evaluation exercises known in literature. We focus in particular on the novel corpora Political debates by describing their development and by comparing them with the benchmarks. Second, Twitter MultiTACOS is applied with different sets of features especially designed for Stance Detection, with a specific focus to exploring and combining both features based on the textual content of the tweet (e.g., style and affective load) and features based on con- textual information that do not emerge directly from the text. Finally, for better highlighting the contribution of the features that most positively affect system perfor- mance in the multilingual setting, a features analysis is provided, together with a quali- tative analysis of the misclassified tweets for each of the observed languages, devoted to reflect on the open challenges. © 2020 Elsevier Ltd. All rights reserved.
1. Introduction
Detecting stance of people towards specific targets is a field of Natural Language Processing (NLP) research that is currently collecting an increasing interest. It has been defined in literature as Stance Detection (SD), that is the task of automatically deter- mining whether the text’s author is in favor, against, or neutral towards a statement or targeted event, person, organization, gov- ernment policy, movement, etc. (Mohammad et al., 2017). Like Sentiment Analysis, also SD has been applied in several domains to discover the reputation of an enterprise, what is the general public thinks of a political reform, if costumers of a fashion brand are happy about the customer service etc. Nevertheless, whereas the aim of Sentiment Analysis is at categorizing texts according to a notion of polarity, i.e. as positive, negative or
*Corresponding author. E-mail addresses: [email protected], [email protected] (M. Lai). 1 The first two authors equally contributed to this work. https://doi.org/10.1016/j.csl.2020.101075 0885-2308/© 2020 Elsevier Ltd. All rights reserved. 2 M. Lai et al. / Computer Speech & Language 63 (2020) 101075 neutral, that of SD consists in classifying texts according to the attitude they express towards a given target of interest. This differ- ence between sentiment polarity and stance can be observed for instance in the following tweeti Target of interest: Climate change is a real concern @RegimeChangeBC @ndnstyl It’s sad to be the last generation that could change but does nothing. #Auspol Polarity: NEGATIVE Stance: FAVOR where the opinion expressed by the user includes a negative polarity contrasting with the stance expressed in favor of the target of interest. This support the idea that SD deserves to be treated as a singular classification task and needs to be distinguished from classical Sentiment Analysis tasks focused on polarity. Several shared tasks already took place for promoting Sentiment Analysis (Bethard et al., 2016; Mohammad et al., 2016b; Nakov et al., 2013; Rosenthal et al., 2015), but only recently SD has been acknowledged as an independent task having its own characteristics, peculiarities and benchmark datasets. The first shared task on SD was indeed held for English at SemEval in 2016, i.e. Task 6 “Detecting Stance in Tweets” (Mohammad et al., 2016b). It consisted in detecting the orientation in favor or against six different targets of interest: “Hillary Clinton”, “Feminist Movement”, “Legalization of Abortion”, “Atheism”, “Donald Trump”, and “Climate Change is a Real Concern”. A more recent evaluation for SD systems was instead proposed at IberEval 2017 for both Cat- alan and Spanish (Taule et al., 2017) where the target was only one, i.e. “Independence of Catalonia”. Not surprisingly in all cases the SD task was based on data extracted from social media, namely Twitter, and about politics and public life topics. On the one hand, social media are contexts where people spontaneously express opinions, desires, complaints, beliefs and outbursts. On the other hand, politics and public life are among the topics mainly discussed by users in social media. In these choices the possible relevance of SD techniques for policy makers and public administrators is also mirrored, e.g. for bet- ter meeting population’s needs and preventing feelings of dissatisfaction and extreme reactions of hostility and anger. A further motivation for collecting texts from social media, in particular Twitter, is that this is a great source of freely available data. Given the increasing interest for the task of automatically detecting stance in political polarized debates on Twitter and the recent development of SD benchmark corpora in several languages, our main focus in this paper is investigating the portability of tools performing SD in a multilingual setting. The languages that we work with in this research are five, namely: English, Spanish, French, Catalan and Italian but a manifold of corpora in other languages is available to the scientific community: Arabic (Magdy et al., 2016), Chinese (Yuan et al., 2019), Russian (Vychegzhanin and Kotelnikov, 2019), and Turkish (Ku€c¸ uk€ and Can, 2019). Our starting point is a model, called MultiTACOS, for addressing SD as a classification problem, which exploits a multifaceted set of features especially designed for identifying stance in Twitter. In particular, we are interested in exploring and combining features based on the textual content of the tweet, such as structural, stylistic, and affective features, but also features based on contextual information that do not emerge directly from the text, such as e.g. knowledge about the domain of the political debate or about the user community. Some of these typologies of features have been successfully exploited and evaluated in SD tasks in monolingual settings, but their combination and contribution, and their potential in a multilingual setting have never been stud- ied. We are in particular interested in finding answers for research questions about: